Skip to contentSkip to conversation

Don't obey the kernel block device max transfer len / max segments for raw block devices

PATCH Applied Maxim Levitsky 4 participants 6 messages Jul 4, 2019 – Jul 12, 2019 Block layer

lore
Share & tools
lore
#0 lore raw

v2 cover Don't obey the kernel block device max transfer len / max segments for raw block devices

Linux block devices, even in O_DIRECT mode don't have any user visible
limit on transfer size / number of segments, which underlying kernel block device can have. The kernel block layer takes care of enforcing these limits by splitting the bios.

By limiting the transfer sizes, we force qemu to do the splitting itself which introduces various overheads.
It is especially visible in nbd server, where the low max transfer size of the underlying device forces us to advertise this over NBD, thus increasing the traffic overhead in case of image conversion which benefits from large blocks.

More information can be found here:
https://bugzilla.redhat.com/show_bug.cgi?id=1647104

Tested this with qemu-img convert over nbd and natively and to my surprise, even native IO performance improved a bit.

(The device on which it was tested is Intel Optane DC P4800X, which has 128k max transfer size reported by the kernel)

The benchmark:

Images were created using:

Sparse image: qemu-img create -f qcow2 /dev/nvme0n1p3 1G / 10G / 100G
Allocated image: qemu-img create -f qcow2 /dev/nvme0n1p3 -o preallocation=metadata 1G / 10G / 100G

The test was:

echo "convert native:"
rm -rf /dev/shm/disk.img
time qemu-img convert -p -f qcow2 -O raw -T none $FILE /dev/shm/disk.img > /dev/zero

echo "convert via nbd:"
qemu-nbd -k /tmp/nbd.sock -v -f qcow2 $FILE -x export --cache=none --aio=native --fork
rm -rf /dev/shm/disk.img
time qemu-img convert -p -f raw -O raw nbd:unix:/tmp/nbd.sock:exportname=export /dev/shm/disk.img > /dev/zero

The results:

=========================================
1G sparse image:
native:

	before: 0.027s
	after: 0.027s
 nbd:
	before: 0.287s
	after: 0.035s

=========================================
100G sparse image:
native:

	before: 0.028s
	after: 0.028s
 nbd:
	before: 23.796s
	after: 0.109s

=========================================
1G preallocated image:
native:

       before: 0.454s
       after: 0.427s
 nbd:
       before: 0.649s
       after: 0.546s

The block limits of max transfer size/max segment size are retained
for the SCSI passthrough because in this case the kernel passes the userspace request directly to the kernel scsi driver, bypassing the block layer, and thus there is no code to split such requests.

Fam, since you was the original author of the code that added these limits, could you share your opinion on that? What was the reason besides SCSI passthrough?

V2:

*  Manually tested to not break the scsi passthrough with a nested VM
*  As Eric suggested, refactored the area around the fstat.
*  Spelling/grammar fixes

Best regards,
Maxim Levitsky

Maxim Levitsky (1):
raw-posix.c - use max transfer length / max segement count only for
SCSI passthrough

 block/file-posix.c | 54 ++++++++++++++++++++++++----------------------
 1 file changed, 28 insertions(+), 26 deletions(-)
Signature · 1 line
2.17.2
#1 lore raw

PATCH v2 1/1 raw-posix.c - use max transfer length / max segement count only for SCSI passthrough

Regular kernel block devices (/dev/sda*, /dev/nvme*, etc) don't have max segment size/max segment count hardware requirements exposed to the userspace, but rather the kernel block layer
takes care to split the incoming requests that
violate these requirements.

Allowing the kernel to do the splitting allows qemu to avoid various overheads that arise otherwise from this.

This is especially visible in nbd server,
exposing as a raw file, a mostly empty qcow2 image over the net. In this case most of the reads by the remote user
won't even hit the underlying kernel block device,
and therefore most of the overhead will be in the
nbd traffic which increases significantly with lower max transfer size.

In addition to that even for local block device access the peformance improves a bit due to less traffic between qemu and the kernel when large transfer sizes are used (e.g for image conversion)

More info can be found at:
https://bugzilla.redhat.com/show_bug.cgi?id=1647104

---
 block/file-posix.c | 54 ++++++++++++++++++++++++----------------------
 1 file changed, 28 insertions(+), 26 deletions(-)
block/file-posix.c +28 -26
index ab05b51a66..4479cc7ab4 100644
@@ -1038,15 +1038,13 @@ static void raw_reopen_abort(BDRVReopenState *state)
     s->reopen_state = NULL;
 }
 
-static int hdev_get_max_transfer_length(BlockDriverState *bs, int fd)
+static int sg_get_max_transfer_length(int fd)
 {
 #ifdef BLKSECTGET
     int max_bytes = 0;
-    short max_sectors = 0;
-    if (bs->sg && ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
+
+    if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
         return max_bytes;
-    } else if (!bs->sg && ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
-        return max_sectors << BDRV_SECTOR_BITS;
     } else {
         return -errno;
     }
@@ -1055,25 +1053,31 @@ static int hdev_get_max_transfer_length(BlockDriverState *bs, int fd)
 #endif
 }
 
-static int hdev_get_max_segments(const struct stat *st)
+static int sg_get_max_segments(int fd)
 {
 #ifdef CONFIG_LINUX
     char buf[32];
     const char *end;
-    char *sysfspath;
+    char *sysfspath = NULL;
     int ret;
-    int fd = -1;
+    int sysfd = -1;
     long max_segments;
+    struct stat st;
+
+    if (fstat(fd, &st)) {
+        ret = -errno;
+        goto out;
+    }
 
     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/max_segments",
-                                major(st->st_rdev), minor(st->st_rdev));
-    fd = open(sysfspath, O_RDONLY);
-    if (fd == -1) {
+                                major(st.st_rdev), minor(st.st_rdev));
+    sysfd = open(sysfspath, O_RDONLY);
+    if (sysfd == -1) {
         ret = -errno;
         goto out;
     }
     do {
-        ret = read(fd, buf, sizeof(buf) - 1);
+        ret = read(sysfd, buf, sizeof(buf) - 1);
     } while (ret == -1 && errno == EINTR);
     if (ret < 0) {
         ret = -errno;
@@ -1090,8 +1094,8 @@ static int hdev_get_max_segments(const struct stat *st)
     }
 
 out:
-    if (fd != -1) {
-        close(fd);
+    if (sysfd != -1) {
+        close(sysfd);
     }
     g_free(sysfspath);
     return ret;
@@ -1103,19 +1107,17 @@ out:
 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
 {
     BDRVRawState *s = bs->opaque;
-    struct stat st;
 
-    if (!fstat(s->fd, &st)) {
-        if (S_ISBLK(st.st_mode) || S_ISCHR(st.st_mode)) {
-            int ret = hdev_get_max_transfer_length(bs, s->fd);
-            if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
-                bs->bl.max_transfer = pow2floor(ret);
-            }
-            ret = hdev_get_max_segments(&st);
-            if (ret > 0) {
-                bs->bl.max_transfer = MIN(bs->bl.max_transfer,
-                                          ret * getpagesize());
-            }
+    if (bs->sg) {
+        int ret = sg_get_max_transfer_length(s->fd);
+
+        if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
+            bs->bl.max_transfer = pow2floor(ret);
+        }
+
+        ret = sg_get_max_segments(s->fd);
+        if (ret > 0) {
+            bs->bl.max_transfer = MIN(bs->bl.max_transfer, ret * getpagesize());
         }
     }
 
Signature · 1 line
2.17.2
#2 lore raw

Re: [Qemu-devel] [PATCH v2 1/1] raw-posix.c - use max transfer length / max segement count only for SCSI passthrough

Quoted 124 lines · Maxim Levitsky

Regular kernel block devices (/dev/sda*, /dev/nvme*, etc) don't have max segment size/max segment count hardware requirements exposed to the userspace, but rather the kernel block layer
takes care to split the incoming requests that
violate these requirements.

Allowing the kernel to do the splitting allows qemu to avoid various overheads that arise otherwise from this.

This is especially visible in nbd server,
exposing as a raw file, a mostly empty qcow2 image over the net. In this case most of the reads by the remote user
won't even hit the underlying kernel block device,
and therefore most of the overhead will be in the
nbd traffic which increases significantly with lower max transfer size.

In addition to that even for local block device access the peformance improves a bit due to less traffic between qemu and the kernel when large transfer sizes are used (e.g for image conversion)

More info can be found at:
https://bugzilla.redhat.com/show_bug.cgi?id=1647104

---
 block/file-posix.c | 54 ++++++++++++++++++++++++----------------------
 1 file changed, 28 insertions(+), 26 deletions(-)
block/file-posix.c +21 -26
index ab05b51a66..4479cc7ab4 100644
@@ -1038,15 +1038,13 @@ static void raw_reopen_abort(BDRVReopenState *state)
     s->reopen_state = NULL;
 }
 
-static int hdev_get_max_transfer_length(BlockDriverState *bs, int fd)
+static int sg_get_max_transfer_length(int fd)
 {
 #ifdef BLKSECTGET
     int max_bytes = 0;
-    short max_sectors = 0;
-    if (bs->sg && ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
+
+    if (ioctl(fd, BLKSECTGET, &max_bytes) == 0) {
         return max_bytes;
-    } else if (!bs->sg && ioctl(fd, BLKSECTGET, &max_sectors) == 0) {
-        return max_sectors << BDRV_SECTOR_BITS;
     } else {
         return -errno;
     }
@@ -1055,25 +1053,31 @@ static int hdev_get_max_transfer_length(BlockDriverState *bs, int fd)
 #endif
 }
 
-static int hdev_get_max_segments(const struct stat *st)
+static int sg_get_max_segments(int fd)
 {
 #ifdef CONFIG_LINUX
     char buf[32];
     const char *end;
-    char *sysfspath;
+    char *sysfspath = NULL;
     int ret;
-    int fd = -1;
+    int sysfd = -1;
     long max_segments;
+    struct stat st;
+
+    if (fstat(fd, &st)) {
+        ret = -errno;
+        goto out;
+    }
 
     sysfspath = g_strdup_printf("/sys/dev/block/%u:%u/queue/max_segments",
-                                major(st->st_rdev), minor(st->st_rdev));
-    fd = open(sysfspath, O_RDONLY);
-    if (fd == -1) {
+                                major(st.st_rdev), minor(st.st_rdev));
+    sysfd = open(sysfspath, O_RDONLY);
+    if (sysfd == -1) {
         ret = -errno;
         goto out;
     }
     do {
-        ret = read(fd, buf, sizeof(buf) - 1);
+        ret = read(sysfd, buf, sizeof(buf) - 1);
     } while (ret == -1 && errno == EINTR);
     if (ret < 0) {
         ret = -errno;
@@ -1090,8 +1094,8 @@ static int hdev_get_max_segments(const struct stat *st)
     }
 
 out:
-    if (fd != -1) {
-        close(fd);
+    if (sysfd != -1) {
+        close(sysfd);
     }
     g_free(sysfspath);
     return ret;
@@ -1103,19 +1107,17 @@ out:
 static void raw_refresh_limits(BlockDriverState *bs, Error **errp)
 {
     BDRVRawState *s = bs->opaque;
-    struct stat st;
 
-    if (!fstat(s->fd, &st)) {
-        if (S_ISBLK(st.st_mode) || S_ISCHR(st.st_mode)) {
-            int ret = hdev_get_max_transfer_length(bs, s->fd);
-            if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
-                bs->bl.max_transfer = pow2floor(ret);
-            }
-            ret = hdev_get_max_segments(&st);
-            if (ret > 0) {
-                bs->bl.max_transfer = MIN(bs->bl.max_transfer,
-                                          ret * getpagesize());
-            }
+    if (bs->sg) {
+        int ret = sg_get_max_transfer_length(s->fd);
+

block/file-posix.c out: · line 1123

Quoted 10 lines
block/file-posix.c +7 -0
+        if (ret > 0 && ret <= BDRV_REQUEST_MAX_BYTES) {
+            bs->bl.max_transfer = pow2floor(ret);
+        }
+
+        ret = sg_get_max_segments(s->fd);
+        if (ret > 0) {
+            bs->bl.max_transfer = MIN(bs->bl.max_transfer, ret * getpagesize());
         }
     }
 

Ping.

Best regards,
Maxim Levitsky

Reviewed-by on the v2 series
#3 lore raw
In reply to Maxim Levitsky
Quoted 84 lines · Maxim Levitsky

Linux block devices, even in O_DIRECT mode don't have any user visible
limit on transfer size / number of segments, which underlying kernel block device can have. The kernel block layer takes care of enforcing these limits by splitting the bios.

By limiting the transfer sizes, we force qemu to do the splitting itself which introduces various overheads.
It is especially visible in nbd server, where the low max transfer size of the underlying device forces us to advertise this over NBD, thus increasing the traffic overhead in case of image conversion which benefits from large blocks.

More information can be found here:
https://bugzilla.redhat.com/show_bug.cgi?id=1647104

Tested this with qemu-img convert over nbd and natively and to my surprise, even native IO performance improved a bit.

(The device on which it was tested is Intel Optane DC P4800X, which has 128k max transfer size reported by the kernel)

The benchmark:

Images were created using:

Sparse image: qemu-img create -f qcow2 /dev/nvme0n1p3 1G / 10G / 100G
Allocated image: qemu-img create -f qcow2 /dev/nvme0n1p3 -o preallocation=metadata 1G / 10G / 100G

The test was:

echo "convert native:"
rm -rf /dev/shm/disk.img
time qemu-img convert -p -f qcow2 -O raw -T none $FILE /dev/shm/disk.img > /dev/zero

echo "convert via nbd:"
qemu-nbd -k /tmp/nbd.sock -v -f qcow2 $FILE -x export --cache=none --aio=native --fork
rm -rf /dev/shm/disk.img
time qemu-img convert -p -f raw -O raw nbd:unix:/tmp/nbd.sock:exportname=export /dev/shm/disk.img > /dev/zero

The results:

=========================================
1G sparse image:
native:

	before: 0.027s
	after: 0.027s
 nbd:
	before: 0.287s
	after: 0.035s

=========================================
100G sparse image:
native:

	before: 0.028s
	after: 0.028s
 nbd:
	before: 23.796s
	after: 0.109s

=========================================
1G preallocated image:
native:

       before: 0.454s
       after: 0.427s
 nbd:
       before: 0.649s
       after: 0.546s

The block limits of max transfer size/max segment size are retained
for the SCSI passthrough because in this case the kernel passes the userspace request directly to the kernel scsi driver, bypassing the block layer, and thus there is no code to split such requests.

Fam, since you was the original author of the code that added these limits, could you share your opinion on that? What was the reason besides SCSI passthrough?

V2:

*  Manually tested to not break the scsi passthrough with a nested VM
*  As Eric suggested, refactored the area around the fstat.
*  Spelling/grammar fixes

Best regards,
Maxim Levitsky

Quoted 11 lines

Maxim Levitsky (1):
raw-posix.c - use max transfer length / max segement count only for
SCSI passthrough

 block/file-posix.c | 54 ++++++++++++++++++++++++----------------------
 1 file changed, 28 insertions(+), 26 deletions(-)
Signature · 1 line
2.17.2
Reviewed-by on the v2 series
#4 lore raw
In reply to Maxim Levitsky
Quoted 96 lines

Linux block devices, even in O_DIRECT mode don't have any user visible limit on transfer size / number of segments, which underlying kernel block device can have.
The kernel block layer takes care of enforcing these limits by splitting the bios.

By limiting the transfer sizes, we force qemu to do the splitting itself which
introduces various overheads.
It is especially visible in nbd server, where the low max transfer size of the
underlying device forces us to advertise this over NBD, thus increasing the traffic overhead in case of image conversion which benefits from large blocks.

More information can be found here:
https://bugzilla.redhat.com/show_bug.cgi?id=1647104

Tested this with qemu-img convert over nbd and natively and to my surprise,Reviewed-by: Stefano Garzarella <[email protected]> even native IO performance improved a bit.

(The device on which it was tested is Intel Optane DC P4800X, which has 128k max transfer size reported by the kernel)

The benchmark:

Images were created using:

Sparse image: qemu-img create -f qcow2 /dev/nvme0n1p3 1G / 10G / 100G Allocated image: qemu-img create -f qcow2 /dev/nvme0n1p3 -o preallocation=metadata 1G / 10G / 100G

The test was:

echo "convert native:"
rm -rf /dev/shm/disk.img
time qemu-img convert -p -f qcow2 -O raw -T none $FILE /dev/shm/disk.img > /dev/zero

echo "convert via nbd:"
qemu-nbd -k /tmp/nbd.sock -v -f qcow2 $FILE -x export --cache=none --aio=native --fork
rm -rf /dev/shm/disk.img
time qemu-img convert -p -f raw -O raw nbd:unix:/tmp/nbd.sock:exportname=export /dev/shm/disk.img > /dev/zero

The results:

=========================================
1G sparse image:
native:

	before: 0.027s
	after: 0.027s
 nbd:
	before: 0.287s
	after: 0.035s

=========================================
100G sparse image:
native:

	before: 0.028s
	after: 0.028s
 nbd:
	before: 23.796s
	after: 0.109s

=========================================
1G preallocated image:
native:

       before: 0.454s
       after: 0.427s
 nbd:
       before: 0.649s
       after: 0.546s

The block limits of max transfer size/max segment size are retained for the SCSI passthrough because in this case the kernel passes the userspace request
directly to the kernel scsi driver, bypassing the block layer, and thus there is no code to split
such requests.

Fam, since you was the original author of the code that added these limits, could you share your opinion on that? What was the reason besides SCSI passthrough?

V2:

*  Manually tested to not break the scsi passthrough with a nested VM
*  As Eric suggested, refactored the area around the fstat.
*  Spelling/grammar fixes

Best regards,
Maxim Levitsky

Quoted 8 lines

Maxim Levitsky (1):
raw-posix.c - use max transfer length / max segement count only for
SCSI passthrough

 block/file-posix.c | 54 ++++++++++++++++++++++++----------------------
 1 file changed, 28 insertions(+), 26 deletions(-)

--

I am not familiar with SCSI passthrough special case. But overall looks good to me.

Feel free to add:

Quoted 4 lines

2.17.2

Applied
#5 lore raw
In reply to Maxim Levitsky

Am 04.07.2019 um 14:43 hat Maxim Levitsky geschrieben:

Quoted 10 lines

Linux block devices, even in O_DIRECT mode don't have any user visible
limit on transfer size / number of segments, which underlying kernel block device can have. The kernel block layer takes care of enforcing these limits by splitting the bios.

By limiting the transfer sizes, we force qemu to do the splitting itself which introduces various overheads.
It is especially visible in nbd server, where the low max transfer size of the underlying device forces us to advertise this over NBD, thus increasing the traffic overhead in case of image conversion which benefits from large blocks.

Quoted 5 lines

More information can be found here:
https://bugzilla.redhat.com/show_bug.cgi?id=1647104

Tested this with qemu-img convert over nbd and natively and to my surprise, even native IO performance improved a bit.

Thanks, applied to the block branch.

Kevin